Lock-in: Research Report

📋Page Status

Quality:3 (Stub)⚠️

Words:4.2k

Backlinks:9

Structure:

📊 17📈 0🔗 0📚 69•18%Score: 10/15

Executive Summary

Finding	Key Data	Implication
Value lock-in risk identified	MacAskill warns AGI could “lock in values indefinitely”	Current moral blind spots may become permanent
Surveillance infrastructure spreading	China-sourced AI surveillance in 80+ countries; 34% global market share	Political system lock-in already occurring
Economic concentration extreme	Nvidia 92% GPU market share; $90B+ AI infrastructure investments	Monopoly power creates technological lock-in
Value embedding operational	Constitutional AI explicitly trains values into models	Embedded values may resist modification
Infrastructure path dependence	24-month lead time for AI capacity; 15-20 year data center lifespans	Early infrastructure choices constrain future options
Irreversibility timeline	5-20 years before lock-in becomes permanent	Intervention window narrowing rapidly

---

Research Summary

Lock-in refers to the permanent entrenchment of values, systems, or power structures in ways that become extremely difficult or impossible to reverse. In the AI context, this represents a pathway to existential catastrophe where early decisions about development, deployment, or governance become irreversibly embedded in future systems. Unlike traditional technologies where course correction remains possible, advanced AI could create enforcement mechanisms so powerful that alternative paths become permanently inaccessible.

Three primary lock-in mechanisms are documented. Value lock-in occurs through training processes like Constitutional AI that explicitly embed ethical principles into models, raising questions about whose values get embedded and whether they can be changed. Political lock-in manifests through AI surveillance deployed in over 80 countries, with Chinese firms controlling 34% of the global surveillance market and lock-in effects making supplier changes prohibitively expensive. Economic lock-in emerges from extreme market concentration—Nvidia holds 92% of data center GPU market share, while training costs exceeding $100M create insurmountable barriers to entry.

Path dependence in AI infrastructure creates additional irreversibility. Organizations must plan AI capacity 24 months in advance or risk being locked out entirely. Data centers optimized for current training workloads face 15-20 year lifespans but may become obsolete if inference patterns shift. Energy infrastructure choices—fossil fuel versus renewable pathways—could lock in decades of climate impact. Researchers including MacAskill and Bostrom warn that technological advances, particularly AGI, could enable those in power to lock in their values indefinitely, making current decisions potentially the most consequential in human history. The 5-20 year timeline before irreversibility suggests each intervention year is more valuable than the next.

---

Background

Lock-in represents a unique category of existential risk where the permanence of outcomes, rather than their immediate severity, defines the threat. As Toby Ord articulates in The Precipice, “dystopian lock-in” could be as serious as extinction—both permanently curtail humanity’s potential, but lock-in preserves awareness of what was lost.

Why This Matters for AI Safety

Lock-in transforms temporary problems into permanent ones. Even recoverable AI failures become irreversible if they occur during a critical window when lock-in becomes possible. This makes understanding lock-in mechanisms essential for effective AI governance.

The concept gained prominence through William MacAskill’s What We Owe the Future (2022), which identifies AI as a key enabler of value lock-in. MacAskill argues we may be living through humanity’s “time of perils”—a period when technological capabilities enable permanent entrenchment of current values before humanity has time to evolve them. Historical examples like slavery demonstrate that widely accepted values can later prove deeply wrong, making premature lock-in catastrophic.

The No-Turning-Back Point

Unlike conventional risks where recovery remains possible, lock-in by definition prevents course correction. Once achieved, enforcement mechanisms maintain the locked-in state regardless of whether it serves human flourishing. This makes prevention the only viable strategy.

Recent research distinguishes “decisive” from “accumulative” existential risks. Lock-in primarily operates through accumulative mechanisms—gradual entrenchment over years to decades rather than sudden catastrophic events. This temporal dimension makes political mobilization difficult, as there’s no clear “stop moment” to rally around.

---

Key Findings

Value Lock-In Through AI Training

Modern AI training explicitly embeds values and objectives into systems in ways that may resist modification. Constitutional AI, developed by Anthropic, trains models to follow a constitution of principles curated by employees, drawing from sources including the UN Universal Declaration of Human Rights and Apple’s terms of service.

Value Embedding Mechanics

Constitutional AI uses a two-stage process: first training models to critique and revise their responses using constitutional principles, then training via reinforcement learning using AI-generated feedback based on those principles. This embeds values deeply into model behavior.

Value Source	Implementation	Lock-in Concern
UN Declaration of Human Rights	Principles like “support freedom, equality and brotherhood”	Western liberal values may not represent global consensus
Corporate terms of service	Apple’s ToS influences model behavior	Commercial interests shape public AI systems
Anthropic employee judgment	Internal curation of principles	Small group determines values for millions of users
Training data distribution	Reflects English-language, Western internet	Cultural biases may become permanent

The irreversibility concern stems from training economics. Advanced models cost over $100M to train, with costs doubling approximately every six months. Once values are embedded, modification requires expensive retraining that may not fully reverse earlier value embedding. As researchers note, “AI systems might resist subsequent attempts to change their goals”—a superintelligent system would likely succeed in out-maneuvering operators attempting to reprogram it.

Moral Progress at Risk

If humanity still has moral blind spots similar to slavery in the past, AI might irreversibly entrench them, preventing moral progress. AI could also be used to spread and preserve the values of whoever develops it, regardless of whether those values serve humanity’s long-term flourishing.

Political System Lock-In via Surveillance

AI surveillance infrastructure is creating conditions for permanent political entrenchment. According to the Carnegie Endowment for International Peace, PRC-sourced AI surveillance solutions have diffused to over 80 authoritarian and democratic countries worldwide. As of 2024, Hikvision and Dahua jointly control approximately 34% of the global surveillance camera market.

Surveillance Metric	Data	Source	Lock-in Mechanism
Countries using Chinese AI surveillance	80+	Carnegie Endowment	Global infrastructure dependency
Global surveillance market share	34% (Hikvision + Dahua)	Carnegie Endowment	Vendor lock-in through incompatibility
China’s AI-powered cameras	200M+	MERICS	Domestic enforcement capability
Social Credit restrictions	23M+ flight bans, 5.5M+ train bans	Multiple sources	Behavioral control at scale

The “lock-in effect” operates through technical incompatibility and switching costs. Systems from different companies are not interoperable, making supplier changes prohibitively expensive. Countries that have come to rely on Chinese surveillance technology become dependent on it, according to Carnegie researchers.

Democracy Under Pressure

Research from the Journal of Democracy documents how AI has become a potent tool for authoritarian control through mass surveillance, facial recognition, predictive policing, and electoral manipulation. Unlike traditional repression tools, AI enables more sophisticated manipulation of information and public perceptions.

China’s domestic implementation provides a proof-of-concept. The Chinese Communist Party has deployed vast networks of AI-driven cameras capable of identifying individuals in real time, making it nearly impossible for activists to operate anonymously. Through the Digital Silk Road initiative, China has become an exporter of digital authoritarianism, with instances observable in Bangladesh, Colombia, Ethiopia, Guatemala, the Philippines, and Thailand.

Economic Lock-In Through Market Concentration

AI market concentration has reached levels that threaten competitive alternatives. Research published in Economic Policy concludes that without clear antitrust rules and regulatory actions, market concentration in generative AI could lead to systemic risks and stark inequality.

Company/Sector	Market Share/Position	Lock-in Mechanism	Implication
Nvidia (GPUs)	92% data center market share	Hardware bottleneck	Single point of failure for AI development
Cloud (Big 3)	66-70% AWS+Azure+Google	Infrastructure integration, data gravity	Dependency on incumbent platforms
Surveillance	34% Hikvision+Dahua	Hardware lock-in, data formats	Political implications of vendor dependency
Training Costs	$100M+ for frontier models	Insurmountable barriers to entry	Only largest firms can compete

The 'Winner-Takes-All' Scenario

Nvidia’s dominance raises concerns about a “winner-takes-all” scenario. The AI industry is consolidating into “the Blob”—a powerful entity formed by intertwined tech giants through massive investments and partnerships, effectively creating monopolies in critical AI infrastructure.

Vertical integration compounds concentration risks. The same companies that manufacture essential chips (Nvidia), provide cloud computing (Amazon, Google, Microsoft), and collect training data (Meta, Google) are also developing the most important AI models. Researchers warn that generative AI firms face strong incentives to integrate vertically with providers of AI building blocks, further consolidating control.

The “context flywheel”—a rich, structured user and project data layer—drives up switching costs, creating lock-in effects that trap accumulated data within platforms. According to a January 2025 FTC report, partnerships like Google-Anthropic and Microsoft-OpenAI risk “locking in the market dominance of large incumbent technology firms.”

Technological Path Dependence in Infrastructure

AI infrastructure decisions create decades-long path dependencies. Industry analysis from S&P Global indicates that with vacancy rates in primary data center markets at record lows, organizations must plan IT infrastructure needs at least 24 months in advance or risk being locked out of capacity required to scale AI.

Infrastructure Dimension	Lock-in Dynamic	Timeline	Reversibility
Data center capacity	Must reserve 24 months ahead	2-year planning horizon	Low—capacity constraints limit alternatives
Training vs. inference optimization	Infrastructure optimized for training may become obsolete	15-20 year facility lifespan	Very low—$90B+ investments
Energy pathway	Fossil fuel vs. renewable choices	Decades of operation	Low—infrastructure replacement costs
Supply chain	TSMC 87% of 5nm+ smartphone SoCs	Multi-year fabrication lead times	Very low—geopolitical concentration risk

The Overbuilding Risk

Analysis suggests the industry may be transitioning from compute-intensive “model training” to less demanding “inference” phases. Data centers optimized for training workloads could find themselves overbuilt and underutilized—a $90B mismatch with 15-20 year design life consequences.

The scale of investment amplifies lock-in. In late 2024, Alphabet announced $40 billion for AI infrastructure while Anthropic committed $50 billion for new data centers. Grid Strategies found that nationally, the utility industry is planning for about 50% more data center demand than the tech industry is projecting—a potential overbuilding that locks in energy infrastructure choices for decades.

Taiwan Semiconductor Manufacturing (TSMC) stands at the epicenter. As the world’s leading foundry, TSMC will likely lead smartphone SoC shipments with 87% share in 5nm and below nodes, expected to grow to 89% by 2028. Constrained supply of leading-edge fabrication has enabled a small group of suppliers to capture majority market share, creating geopolitical risk and single points of failure.

Irreversibility of AI Deployment

Once highly capable AI systems are deployed, correcting misalignment becomes extremely difficult. Research from arXiv argues that the recursive failure to assess certain alignment models “is not just a sociological oversight but a structural attractor, mirroring the very risks of misalignment we aim to avoid in AGI.” Without adopting models of epistemic correction, “we may be on a predictable path toward irreversible misalignment.”

The Option Value of Caution

Researchers note that due to the irreversibility of existential risk, pausing AI development may preserve option value when there is uncertainty about how risky AI is. Once lock-in occurs, the option to choose differently is permanently foreclosed.

Irreversibility Mechanism	Evidence	Timeline	Prevention Window
Value embedding resistance	Models may resist goal changes	Post-deployment	Pre-training only
Economic sunk costs	$100M+ training investments	Years to decades	During development
Infrastructure dependencies	Critical systems require AI	5-15 years	Current period
Political entrenchment	Surveillance enables permanent control	10-30 years	Next decade

MacAskill’s Framework: Value Lock-In as Existential Risk

William MacAskill’s What We Owe the Future provides the most comprehensive treatment of lock-in as existential risk. MacAskill warns of potential value lock-in—“an event that causes a single value system to persist for an extremely long time”—which may result from technological advances, particularly AGI development.

Three mechanisms enable AI-driven lock-in:

1. AGI agents with aligned goals: People may create AGI agents with goals closely aligned with their own that act on their behalf indefinitely
2. Hard-coded objectives: Someone could carefully specify what future they want and ensure the AGI aims to achieve it
3. Mind uploading: People could “upload” themselves, potentially achieving indefinite lifespan while maintaining current values

Why Current Values Matter

As MacAskill notes, if modern-day humanity builds superintelligent machines that use our current values to steer civilization, then anything that affects contemporary values and governance will have outsized long-term impacts. This is the “values lock-in” argument for why changing priorities and moral commitments today could affect the distant future.

Robin Hanson’s analysis emphasizes that MacAskill sees “advanced artificial intelligence” as enabling “those in power to lock in their values indefinitely.” The concern is not merely that bad actors might lock in harmful values, but that even well-intentioned actors might lock in current moral understanding before humanity has time to evolve it.

---

Causal Factors

The following factors influence lock-in probability and severity. These tables are designed to inform future cause-effect diagram creation.

Primary Factors (Strong Influence)

Factor	Direction	Type	Evidence	Confidence
AI Training Costs	↑ Lock-in	cause	$100M+ frontier models; doubling every 6 months	High
Market Concentration	↑ Lock-in	cause	Nvidia 92% GPU share; Big 3 Cloud 66-70%	High
Surveillance Infrastructure	↑ Political Lock-in	intermediate	80+ countries using Chinese AI; 34% market share	High
Value Embedding Methods	↑ Value Lock-in	intermediate	Constitutional AI explicitly trains values	High
Infrastructure Lead Times	↑ Path Dependence	cause	24-month planning horizon; 15-20 year lifespans	High
Supply Chain Concentration	↑ Technological Lock-in	cause	TSMC 87% advanced fabrication	High

Secondary Factors (Medium Influence)

Factor	Direction	Type	Evidence	Confidence
Switching Costs	↑ Lock-in	intermediate	Data gravity; context flywheel effects	Medium
Regulatory Fragmentation	↑ Lock-in	leaf	EU vs China vs US different approaches	Medium
Democratic Input Mechanisms	↓ Lock-in	leaf	Collective Constitutional AI experiments	Medium
Antitrust Enforcement	↓ Economic Lock-in	leaf	DOJ actions; effectiveness uncertain	Medium
Open Source Alternatives	↓ Lock-in	intermediate	Limited viability at frontier scale	Medium
Energy Infrastructure	↑ Path Dependence	cause	Fossil fuel choices lock in decades of emissions	Medium

Minor Factors (Weak Influence)

Factor	Direction	Type	Evidence	Confidence
Public Awareness	↓ Lock-in	leaf	Growing concern but limited actionability	Low
International Coordination	↓ Lock-in	leaf	UN AI Advisory Body; limited enforcement	Low
Transparency Requirements	↓ Lock-in	intermediate	EU AI Act provisions; implementation TBD	Low

---

Lock-In Variants and Manifestations

Interconnected Mechanisms

These lock-in variants are not mutually exclusive. They often co-occur and reinforce each other—economic concentration enables political influence, which enables value embedding, which enables further economic concentration.

Variant	Mechanism	Timeline	Warning Signs	Current Status
Value Lock-in	Training embeds specific moral/political values	10-30 years	Constitutional AI deployment; CCP value mandates	Early stage
Political Lock-in	AI surveillance enables permanent authoritarian control	5-20 years	80+ countries using Chinese surveillance	Actively occurring
Economic Lock-in	Market concentration creates irreversible monopolies	5-15 years	Nvidia 92% share; $100M+ training costs	Rapidly advancing
Technological Lock-in	Infrastructure choices constrain future options	15-30 years	$90B+ data center investments; 24-month lead times	Current window
Epistemic Lock-in	AI-mediated information shapes beliefs permanently	10-25 years	Recommendation algorithm dominance	Early stage

Value Lock-In

Permanent embedding of specific moral, political, or cultural values in AI systems that shape human society. This occurs through:

• Training Data Lock-in: Models trained on Western internet data embed cultural perspectives
• Objective Function Lock-in: Systems optimizing for specific metrics reshape society around those metrics
• Constitutional Lock-in: Explicit value systems embedded during training become permanent features

Researchers note that value lock-in involves “freezing current moral and political perspectives before humanity has time to evolve them.” These risks emerge from broader systemic dynamics rather than requiring any single AI system to behave badly.

Political System Lock-In

AI-enabled permanent entrenchment of particular governments or political systems. Analysis indicates that AI law enforcement tends to undermine democratic government, promote authoritarian drift, and entrench existing authoritarian regimes by reducing structural checks on executive authority.

Journal of Democracy research documents how, through mass surveillance, facial recognition, predictive policing, online harassment, and electoral manipulation, AI has become a potent tool for authoritarian control. Modern technologies promising automated “social management” represent the culmination of a decades-old authoritarian vision: leveraging data to “engineer away dissent.”

Economic Structure Lock-In

AI-enabled economic arrangements that become self-perpetuating and impossible to change through normal market mechanisms. Yale Law & Policy Review analysis warns that concentration in AI markets amplifies the digital divide, with central urban areas and larger economic actors benefiting disproportionately while peripheral regions and SMEs are excluded.

The concern extends beyond consumer harm to democratic governance. Society stands at a critical crossroads: one path leads to AI controlled by a small number of powerful corporations, threatening innovation, privacy, liberty and democracy; the other path leads to a diverse and competitive market where AI serves the broader public interest.

Technological Lock-In

Specific AI architectures or approaches becoming so embedded in global infrastructure that alternatives become impossible. This occurs through:

• Infrastructure Dependencies: AI systems integrated into power grids, financial systems, transportation
• Network Effects: Data advantages and switching costs make dominant platforms unassailable
• Capability Lock-in: Particular architectures achieve advantages making alternatives uncompetitive

Infrastructure analysis indicates the market will generate over $250 billion in 2025, with durable pricing power and high barriers to entry across the infrastructure stack due to constrained supply.

---

Timeline and Intervention Windows

The Narrowing Window

Research consistently points to a 5-20 year timeline before lock-in becomes irreversible. Each year of delay makes intervention more costly and less effective. There is no clear “stop moment”—lock-in accumulates gradually, making political mobilization difficult.

Period	Lock-in Risk	Key Developments	Intervention Options
2024-2026	Low-Medium	Market concentration accelerating; training costs exploding	Antitrust enforcement; regulatory frameworks; open alternatives
2027-2030	Medium	AGI timelines approaching; infrastructure choices locked in	Governance mechanisms; democratic input; transparency requirements
2031-2035	Medium-High	Advanced AI deployment; surveillance global; economic concentration extreme	Value learning systems; international coordination; shutdown capabilities
2036-2045	High	Potential irreversibility; embedded systems; path dependence complete	Very limited; may require radical restructuring
2046+	Very High	Lock-in permanent unless prevented earlier	None—prevention only viable strategy

Expert predictions for 2026 indicate that AI has ceased to be an “emerging” policy issue. Real world harms are accumulating rapidly, putting pressure on lawmakers. The stage is set for important political and legal battles that will define who controls AI, who bears the costs of its harms, and whether democratic governments can keep pace.

---

Open Questions

Key Uncertainties

These questions represent the highest-value areas for follow-up research and intervention development.

Question	Why It Matters	Current State
Can embedded values be modified post-training?	Determines whether value lock-in is reversible	Unknown; $100M+ retraining costs suggest limited reversibility
What level of market concentration enables lock-in?	Defines intervention thresholds for antitrust	Nvidia at 92% GPU share; unclear if already past threshold
Do democratic AI governance mechanisms work?	Determines legitimacy and adaptability of systems	Collective Constitutional AI experiments ongoing; effectiveness TBD
How long until surveillance infrastructure becomes irreversible?	Defines intervention window for political lock-in	80+ countries already using Chinese systems; switching costs high
Can open source alternatives prevent concentration?	Viability of competitive pluralism	Limited success at frontier scale due to compute costs
What triggers irreversibility in value lock-in?	Need to identify point of no return	Theoretical models exist; empirical validation lacking
How do different lock-in types interact?	Economic + political + value lock-in may compound	Likely reinforcing but mechanisms unclear
Is infrastructure overbuilding creating flexibility?	Excess capacity might enable alternatives	Unclear; may lock in wrong architecture if training→inference shift occurs

---

Prevention and Mitigation Strategies

While lock-in by definition prevents recovery once achieved, several strategies may reduce probability or delay onset:

Technological Diversity

Strategy	Mechanism	Status	Effectiveness
Open source models	Prevent monopoly control	Limited at frontier	Medium—cost barriers remain
Research pluralism	Multiple AI approaches	Active globally	Medium—convergence pressure high
Interoperability requirements	Reduce switching costs	EU considering	Unknown
Public cloud infrastructure	Break vendor lock-in	Proposed not implemented	Unknown

Democratic Governance

Strategy	Mechanism	Status	Effectiveness
Collective Constitutional AI	Democratic value input	Experimental	Low—scale unclear
Citizens’ assemblies	Broad stakeholder input	Limited deployment	Low—influence on development unclear
Transparency requirements	Enable oversight	EU AI Act provisions	Medium—enforcement TBD
International coordination	Prevent race to bottom	UN AI Advisory Body	Low—limited enforcement

Antitrust and Competition Policy

Strategy	Mechanism	Status	Effectiveness
Vertical integration restrictions	Prevent concentration	DOJ investigating	Unknown—legal challenges likely
Data portability mandates	Reduce lock-in	Proposed not implemented	Medium—if implemented
Structural separation	Break up incumbents	Proposed for Google	Unknown—political feasibility low
Non-discrimination obligations	Open access to essentials	Under consideration	Medium—enforcement challenges

Technical Safeguards

Strategy	Mechanism	Status	Effectiveness
Value learning systems	Adapt as values evolve	Active research	Unknown—capabilities unclear
Constitutional flexibility	Mechanisms to update values	Theoretical	Unknown—may conflict with stability
Interpretability requirements	Enable human oversight	Active research	Low—advanced systems resist interpretation
Shutdown capabilities	Preserve human control	Limited effectiveness	Low—sophisticated systems may resist

The Prevention-Only Strategy

Given that lock-in by definition prevents reversal, prevention is the only viable strategy. Once values, systems, or power structures become permanently entrenched, recovery becomes impossible by design. This makes understanding and addressing lock-in mechanisms during the current intervention window absolutely critical.

---

Sources

Academic Research

• Hendrycks, D. et al. “X-Risk Analysis for AI Research” - Framework categorizing existential risks including value lock-in, enfeeblement, eroded epistemics
• “The AI Risk Spectrum: From Dangerous Capabilities to Existential Threats” - Value lock-in as freezing moral perspectives before humanity evolves them
• “Introduction to AI Existential Risks” (arXiv, Nov 2025) - Value lock-in mechanisms and survey data on extinction risk
• “Epistemic Closure and Irreversibility of Misalignment” - 2025 paper on structural path to irreversible misalignment
• MacAskill, W. “What We Owe the Future” - Comprehensive treatment of value lock-in as existential risk
• Hanson, R. “MacAskill on Value Lock-In” - Analysis of MacAskill’s lock-in framework

AI Governance and Policy

• TechPolicy.Press “Expert Predictions on AI Policy in 2026” - 2026 outlook on political battles over AI control
• Cloud Security Alliance & Google Cloud “Governance Maturity Study” - Link between governance and AI adoption security

Surveillance and Political Lock-In

• Carnegie Endowment “Can Democracy Survive AI?” - PRC surveillance diffusion to 80+ countries; 34% market share; lock-in effects
• National Endowment for Democracy “Data-Centric Authoritarianism” - China’s role in spreading digital authoritarianism globally
• Journal of Democracy “How Autocrats Weaponize AI” - AI as tool for authoritarian control through surveillance and manipulation
• Journal of Democracy “The Road to Digital Unfreedom” - How AI is reshaping repression
• Yale Review “The Rise of Digital Authoritarianism” - Impacts on democracy and human rights
• Democratic Erosion “AI and Authoritarian Governments” - AI undermining democratic government

Economic Concentration and Monopoly

• Economic Policy “AI Monopolies” - Market concentration could lead to systemic risks and inequality
• TechPolicy.Press “AI Monopolies Are Coming” - FTC report warning on incumbent firm dominance
• Yale Law & Policy Review “Antimonopoly Approach to AI” - Vertical integration and concentration concerns
• Open Markets Institute “Stopping Big Tech from Becoming Big AI” - Report on preventing monopoly control
• AI Frontiers “Open Protocols Can Prevent AI Monopolies” - Model Context Protocol as anti-lock-in tool
• TRENDS Research “Big Tech’s Monopoly of AI” - Threats to fair competition
• WebProNews “AI Giants Form ‘The Blob’” - Consolidation sparking monopoly concerns; Nvidia 92% market share
• The Asset “AI Revolution Cementing Big Tech Monopoly” - How AI revolution strengthens monopolies

Infrastructure and Path Dependence

• S&P Global “AI Infrastructure: Midyear 2025 Update” - 24-month planning horizon; vacancy rate concerns
• Development Corporate “The AI Infrastructure Bubble” - $90B boom could end in bust; training vs inference mismatch
• Cirion Technologies “AI Infrastructure Consolidation 2025” - Whether infrastructure consolidates in 2025
• Data Center Knowledge “AI Infrastructure Revolution” - Predictions for 2026
• IEEFA “Risk of AI-Driven, Overbuilt Infrastructure” - Utility planning exceeds tech industry projections by 50%
• Flexential “State of AI Infrastructure Report 2025” - Market to generate $250B+ in 2025
• INET “The U.S. Is Betting the Economy on ‘Scaling’ AI” - Critique of scaling assumptions
• ScienceDirect “Energy and Climate Alignment of AI Infrastructure” - Energy pathway lock-in concerns
• Sands Capital “Unleashing AI’s Next Wave” - TSMC 87% fabrication market share

Constitutional AI and Value Embedding

• Anthropic “Constitutional AI: Harmlessness from AI Feedback” - Original Constitutional AI methodology
• Anthropic “Claude’s Constitution” - UN Declaration, Apple ToS, employee judgment sources
• Anthropic “Collective Constitutional AI” - Public input approach
• TechCrunch “Anthropic thinks ‘constitutional AI’ is the best way” - Value embedding through training
• Constitutional.ai “Tracking Anthropic’s AI Revolution” - Constitutional AI overview

Existential Risk Frameworks

• Wikipedia “Existential risk from artificial intelligence” - Overview of AI x-risk including lock-in
• Wikipedia “AI alignment” - Irreversibility of advanced AI deployment
• Medium “AI Alignment: The Hidden Challenge” - Humanity’s future depends on alignment success
• EA Forum “The AI Dilemma: Growth vs Existential Risk” - Extension for effective altruists
• EA Forum “My take on What We Owe the Future” - Value lock-in argument analysis
• LSE “Value Alignment Without Institutional Change” - Institutional change necessary for risk mitigation
• Synthese “Current cases of AI misalignment” - Implications for future risks

---

AI Transition Model Context

Model Integration

Lock-in affects multiple factors in the AI Transition Model, particularly Civilizational Competence (governance, adaptability, epistemics), AI Ownership (concentration of control), and Long-term Trajectory (whether outcomes are recoverable or permanent).

Connections to Model Elements

Model Element	Relationship
Civilizational Competence → Governance	Insufficient governance enables lock-in; concentration enables governance capture
Civilizational Competence → Epistemics	AI-mediated information shapes beliefs; epistemic degradation prevents recognition
Civilizational Competence → Adaptability	Lock-in by definition prevents adaptation; flexibility mechanisms critical
AI Ownership → Concentration	Market concentration and surveillance infrastructure are lock-in mechanisms
Misuse Potential → AI Control Concentration	Power concentration creates conditions for political lock-in
Long-term Trajectory	Lock-in determines whether outcomes are temporary or permanent

Lock-in is the defining feature of the Long-term Lock-in scenario—whether values, power, or epistemics become permanently entrenched. This affects Long-term Trajectory more than acute existential risk, as it transforms recoverable problems into permanent ones.